Methylation Preprocessing
Project: Differential Gene
Expression and Pulmonary Function in Vaping LatinX Adolescents
Analyst: Trent Hawkins
Investigator(s): Sunita Sharma, MD
Report generated: September 15, 2022
Analyst: Trent Hawkins
Investigator(s): Sunita Sharma, MD
Report generated: September 15, 2022
Introduction
The purpose of this report is to document quality-control and data transformation procedures followed to obtain methylation values (\(\beta\) and M) for the adolescent vaping study. Much of the code for this report had already been produced by the previous analyst, Cuining Liu. Further steps were taken to dissect the data-pipeline and to support any decisions made moving forward.
Update 09/09/2022
The previous version of this report (“Methylation_QC_2022_09_08.html”) saw no clustering by vape status, the feature of interest for this study. One concern was the absence of within-array normalization. To address this concern, a ‘BMIQ’ normalization was performed in addition to the steps previously presented. ‘noob’ + ‘BMIQ’ has shown to be a reasonable procedure to improve signal sensitivity previously (Liu and Siegmund 2016).
Methods
Methylation samples were taken using the 850K ‘EPIC’ array. All
quality control steps were conducted using the package
SeSAMe ver. 1.15.7. The steps below are not necessarily
presented in chronological order which they were taken. The order of
pre-processing steps follows the openSesame data pipeline
available here
(Zhou et al. 2018).
Sample Removal
Samples will be evaluated for a low mean-intensity to identify potential outliers with low-quality methylation values. Outliers will be defined by samples in the bottom 1%.
Probe Masking
The SeSAMe package uses probe ‘masking’ instead of probe
removal. These terms are interchangeable; however, it should be noted
that ‘masked’ probes are not actually removed from the dataset. They are
only ignored by SeSAMe when completing down-stream
analyses. Probes are masked in two ways:
- Experiment-independent Probe Masking: Probes that are masked due to non-unique mapping or influence by SNPs.
- Experiment-dependent Probe Masking: Probes that are masked based on the detection p-value, which represents the probability of a detection signal being background flourescence. Masks are based on the detection p-value, which represents the probability of a detection signal being background fluorescence (Zhou et al. 2018).
Experiment-independent probe masking is set by a pre-determined list of probes specific to the ‘EPIC’ array. Experiment -dependent probe masking will be determined by a p-value threshold. CpG probes with a detection p-value > 0.05 in at least 10% of samples will be masked for the purpose of this study.
Data Transformations
There are two data transformations within the sesame
pipeline. The first is a dye bias correction. Dye bias refers to bias in
the methylation values due to the performance of the red and green dyes
that create flourescence and are then interpreted as methylation values.
There are several corrections that can be applied to make the red and
green values more comparable. sesame specifically
implements a non-linear dye bias correction.
The next transformation within the openSesame pipeline
is background subtraction. The purpose of background subtraction is to
align the distribution of beta values for Infinium I and II probes in
order to make them more comparable. The default method for
sesame is a normal-exponential deconvolution using
out-of-band probes or ‘noob’ (Triche et al.
2013).
The previous version of this report did not include BMIQ normalization, but this report will to address the concern for insufficient within-array normalization. BMIQ works in conjunction with noob to correct type II probe bias (Liu and Siegmund 2016).
Beta and M-Value Distributions
The product of the SeSAMe data pipeline is a matrix of
beta values. Beta values will be converted to M-values using the logit
transformation
\[M = log_2(\frac{\beta}{1 - \beta})\]
Visualizations will ensure the proper distribution of \(\beta\) and M-values, respectively.
Sample Clustering
Visualizations of clustering by sex, recruitment center, and vape status will help to detect any technical effects that need to be accounted for in downstream analyses. CpG sites chosen for these plots are the top 1000 or 10000 most variable sites.
Results
Of the 51 subjects surveyed for this project, methylation data were collected for 48. Two of the three subjects (SID 105 & 137) for whom there was no transcript data also had no methylation data. One sample (SID 102) included in the RNASeq Analysis known as “Sample 12” also had missing methylation data. This sample was the center of a sensitivity analysis for inclusion in RNASeq analyse (see “sample12_sensitivity_report_2022_mm_dd.html”), and it should be noted that this subject will not be available for comparison. One sample (SID 144) lacked vape status in clinical metadata and was removed after obtaining beta values for all subjects in order to retain the benefits of that sample for normalization purposes. This sample was also excluded from the RNASeq analysis. Downstream analyses will include 47 samples.
| Subject ID | RNASeq ID | Methylation ID |
|---|---|---|
| 102 | Sample12 | No Data |
| 105 | No Data | No Data |
| 137 | No Data | No Data |
Outlier Samples
127 falls in the bottom 1% of mean signal intensity readings.
Figure 1: Sample Quality Outliers
Considering the small sample size in this experiment, it is advisable not to remove 127, but it is worth noting for downstream analyses.
Probe Masking
105,454 CpG sites were removed by non-experimental probe masking and 14,349 were removed by experimental probe masking (detection p-value > 0.05 in \(\ge\) 10% of samples). 119,803 probes were removed in total. 746750 were included for the anlyses presented in this report.
Data Transformations
Figure 2 demonstrates the non-linear dye bias correction in conjunction with ‘noob’ followed by noob + BMIQ for a random subset of samples.
Figure 2: Dye Bias Correction
The inclusion of a dye-bias correction and noob shows improvement over the raw data, but there is still a slight bias towards green which would correspond to unmethylated probes. These plots suggest the the inclusing of the “noob + bmiq” normalization method induces additional bias towards green probes and is visually more biased than the raw data.
Figure 3 demonstrates a combination of the ‘noob’ and ‘BMIQ’ method of background subtraction for three randomly selected samples.
Figure 3: Sample Beta Value Distribution for Infinium I and II Probes
‘noob’ background subtraction in conjunction with ‘BMIQ’ normalization shifted the distributions of probes to have modes closer to 0 and 1 (the expected distribution of beta values). Additionally, it appears that some noise was removed from the unmethylated beta values.
Beta and M-Value Distributions
Overall, \(\beta\) and M-values followed their expected distributions with peaks near 0 (unmethylated sites) and 1 (methylated sites) in the beta distribution.
Figure 4: \(\beta\) and M-Value Distributions
Figure 4 indicates some noise around \(\beta\) = 0.5 or \(M\) = 0. This noise could indicate probes that were not removed by experiment-dependent probe masking. Noise may be reduced at a more stringent threshold for detection p-values (e.g. < 0.01).
Clustering by Sex
Visualization of the samples by median X and Y chromosome intensities helps to identify samples with poor quality or samples whose clinical sex do not match predicted sex based on these values. Figure 5 displays plots of median intensity for X and Y chromosomes color-coded for both clinical and predicted sex.
Figure 5: Sex Verification
Figure 5 demonstrates sample clustering by sex for median X and Y intensities. Clinical sex matches predicted sex for all samples. It should be noted that SID 122 recorded ‘non-binary’ as clinical sex, and sex was inferred by methylation data for other portions of this project.
Sample Clustering
MDS plots were made using the package minfi ver. 1.43.0
using \(\beta\) and M-values
preprocessed using the sesame pipeline. For each feature of
interest, plots were made a) using probes from the whole genome and b)
using only autosomal probes. Using only autosomal probes removes the
inherrent separation by sex to get a better idea of clustering
patterns.
Figure 6: Sample Clustering (Sex)
All Chromosomes (Including X and Y)
Autosomes Only
When retaining the X and Y chromosomes, the samples cluster by sex, as expected. When looking only at the autosomal chromosomes, there is no additional clustering that needs to be accounted for.
Figure 7: Sample Clustering (Vape Status)
The previous version of this report (“Methylation_QC_2022_09_08.html”) saw no clustering by vape status. After the addition of ‘BMIQ’ normalization, there may be some weak clustering by vape status. It is also worthwhile to check for clustering along further MDS coordinates.
Figure 8: Vape Status Clustering (MDS 3 & 4)
Looking at MDS coordinates 3 and 4 for vape status, there is still no obvious clustering pattern. This may indicate technical effects still present in the data that were not handled well by ‘noob + bmiq’ normalization, or it could indicate a lack of biological signal. It is hard to say at this stage, but perhaps other normalization procedures are necessary.
Figure 8: Sample Clustering (Recruitment Center)
There is no visual clustering by recruitment center.
Conclusions
The review of sample quality and clustering patterns returned one subject (SID = 127) that fell in the 1st percentile for mean intensity. There is also a clear sex effect in the data that will be accounted for by including sex as a model covariate. The inclusion of the “noob + BMIQ” normalization method may have induced some weak clustering by vape status, but also induced noticeable bias towards green dye.
References